# RED baseline — no gpt-orchestrator skill

Audience: AI coding agents first.

Run date: 2026-08-11

## Eval 1 — active remote jobs after local wrapper failure

Observed strengths: preserved active conversations; rejected duplicate dispatch; built ledger; reused existing thread; verified and landed incrementally.

Observed failures:

- Omitted account-wide 55–75 second spacing across every new/resume prompt send.
- Preferred manual sidebar reconciliation instead of current `ask-gpt --list`, `--search`, `--list-all`, local registry, and `resume:` evidence ladder.
- Said “Monitor the existing conversations” without pinning notification-driven waiting and no rapid polling.

Verbatim evidence:

> “Reconcile all seven jobs against the ChatGPT sidebar in parallel”

> “Monitor the existing conversations through the supported browser/wrapper resume path”

## Eval 2 — claimed ZIP plus Markdown source over dirty WIP

Observed strengths: treated prose as untrusted; quarantined archive; validated paths/types/limits; preserved WIP; used isolated worktree; ran security and project gates.

Observed failure: permitted reconstruction from inline Markdown, contrary to required real-downloadable-artifact contract and known Markdown transport corruption.

Verbatim evidence:

> “Use printed Markdown only when attachment retrieval is conclusively unavailable and urgency requires reconstruction.”

Required correction: inline source dump is NEVER deliverable fallback. Resume same conversation and demand real downloadable ZIP plus apply/test script.

## Eval 3 — partial, generating, and artifact-ready tasks

Observed strengths: same-thread correction; parallel independent verification; sequential collision-aware integration; no rapid polling; evidence-based states.

Observed failures:

- `DONE` stopped at integrated gates; omitted repository landing plus deploy/install/arm and owner-visible verification.
- Omitted account-wide prompt-send spacing.
- Did not pin artifact retrieval through `--download-attachments` before `ARTIFACT_READY`.

Verbatim evidence:

> “`DONE`: all acceptance criteria and final integrated gates pass.”

## RED contract test

`python3 modules/workstation/claude/tests/gpt-orchestrator-skill.test.py` failed as expected:

```text
AssertionError: missing global skill source: .../skills/gpt-orchestrator/SKILL.md
```